Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a zero-shot detector for AI-generated images that scores inputs with curvature, gradient and bias quantities estimated from a pre-trained Stable Diffusion score function, and extends it to few-shot via a mixture-of-experts with a CLIP-based few-shot detector.
Problem
Detecting AI-generated images in zero-shot and few-shot regimes, where existing methods lack theoretical grounding and generalize poorly to unseen generative techniques.
Approach
Derives a fixed-point score-function criterion from a pre-trained diffusion model that combines total-variation curvature, gradient magnitude and a statistical bias term, estimated via spherical perturbations of the input image; generated images are expected to sit near stable local maxima of the learned log-probability manifold.
Key results
- Zero-shot average AUC 0.835 and AP 0.832 across 20 generative techniques, versus 0.439 AUC for RIGID and 0.444 for AEROBLADE.
- Zero-shot average accuracy 0.741, compared with 0.555 for RIGID and 0.464 for AEROBLADE.
- Sensitivity: AUC drops to 0.79 under JPEG compression and 0.822 under Gaussian blur.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the manifold-bias criterion instead of RIGID or AEROBLADE | Zero-shot image detection · 20 generation techniques | AUC: RIGID 0.439; AEROBLADE 0.444; Ours 0.835 | Helped |
| Use the manifold-bias criterion instead of RIGID or AEROBLADE | Zero-shot image detection · 20 generation techniques | AP: RIGID 0.519; AEROBLADE 0.492; Ours 0.832 | Helped |
| Use the manifold-bias criterion instead of RIGID or AEROBLADE | Zero-shot image detection · 20 generation techniques | Accuracy: RIGID 0.555; AEROBLADE 0.464; Ours 0.741 | Helped |
| Increase the number of spherical perturbations (ablation) | Ours, Stable Diffusion 1.4 · 20 generation techniques | AUC: 0.828 -> 0.829 -> 0.8308 -> 0.833 -> 0.835 | Helped |
| Increase spherical noise level α from 0.01 (ablation) | Ours, Stable Diffusion 1.4 · 20 generation techniques | AUC: 0.835 -> 0.82; 0.835 -> 0.815 | Hurt |
| Apply JPEG compression (ablation) | Ours, Stable Diffusion 1.4 · 20 generation techniques | AUC: 0.835 -> 0.79 | Hurt |
| Apply Gaussian blur (ablation) | Ours, Stable Diffusion 1.4 · 20 generation techniques | AUC: 0.835 -> 0.822 | Hurt |
| Replace SD v1.4 with SD v2 base (ablation) | Ours · 20 generation techniques | AUC: 0.835 -> 0.831 | Hurt |
| Replace SD v1.4 with Kandinsky 2.1 (ablation) | Ours · 20 generation techniques | AUC: 0.835 -> 0.826 | Hurt |
Surprising findings
- Competitor methods show inconsistent real-vs-fake criterion ordering across techniques (e.g., RIGID good on Wukong but near-zero AUC on Glide 50), so a global threshold fails for them.
- The method detects images from generators other than the analyzed Stable Diffusion model, which the authors note lacks a comprehensive theoretical explanation.
Limitations
- No comprehensive theory explains cross-model generalization to generators other than the analyzed diffusion model.
- Derivations rely on approximations (concentration of measure, high-dimensional interchangeability of noise and perturbations) that are not rigorously validated.
- Sensitive to image corruptions such as JPEG compression (AUC drop to 0.79).
Experimental setup
Datasets: CNNSpot Universal Fake Detect GenImage LSUN MSCOCO ImageNet LAION
Benchmarks: CNNSpot Universal Fake Detect GenImage
Metrics: AUC AP Accuracy
Baselines: RIGID AEROBLADE
Builds on: DetectGPT-style curvature criterion AEROBLADE RIGID CLIP-based few-shot detection
Closest papers by meaning
- Detecting Generated Images by Fitting Natural Image Distributions (NeurIPS 2025 · similarity 0.91)
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · similarity 0.87)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.86)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.86)
- Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images (CVPR 2025 · similarity 0.86)
- Categorical Flow Matching on Statistical Manifolds (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.